Capabilities

Months since a US model first matched China's best, by capability area. There is no single overall number: the three overall measures disagree.

Open capabilities

Compute

China's AI chip stock and the best chips it actually has in use, against seven US companies, each measured with the same rule.

Open compute

Cost

The cheapest model from each country that reaches a given score on the Artificial Analysis Intelligence Index, at list price per million tokens.

Open cost

Also here: a timeline of export controls and Chinese chip milestones with where each measure stood at the time, and how every number is calculated, with the data still to check.

Where China is now on the US calendar (median of the area's benchmarks) Low confidence: fewer than 3 benchmarks, or benchmarks that disagree on direction A published estimate, shown where this data has fewer than 2 benchmarks One benchmark Median under the other source setting, where it differs

Show

Events are marked for reference only. A change in the lag after an export control or a chip milestone does not show that one caused the other. The lag also moves for unrelated reasons: models are released in bursts, and benchmarks enter and leave the calculation as they gain coverage or saturate.

The shaded years have limited Chinese coverage. Before 2025 Epoch scored only open-weight Chinese models, and Arena and Artificial Analysis rated few Chinese models of any kind, so lags shown for 2023 and 2024 are less reliable than later ones.

Every area since 2023

Median lag in months at the end of each quarter, using only the benchmarks that had enough recent Chinese scores at the time. The last column is the current quarter to date. Blank means none qualified. Select a cell to open that area.

Faded: low confidence (fewer than 3 benchmarks, or benchmarks that disagree on direction) Shaded years: before 2025 only open-weight Chinese models were widely scored, so the early lag is less reliable

Published estimates beside this data

Grey is what the publisher reported. Blue is the same quantity recomputed here for the same date, where this data can measure it. Methods and model sets differ, so expect them to differ. Hover a row for detail.

Compute

How much AI computing power each side has, how good the chips it actually uses are, and how large its biggest disclosed training runs are. Chip stock and deployed chips use the same rule as the capability lags: when did the US side first reach China's current level?

AI chip stock

Cumulative AI chips acquired since the start of 2022, in H100-equivalents, on a log scale. Estimates by Epoch AI, through its last complete quarter.

What "China" means here. It is Epoch's estimate of chips sold to companies based in China (Nvidia's H800, H20 and others, AMD's MI308X) plus Chinese-designed chips (Huawei Ascend, Cambricon). It does not mean government ownership. It excludes smuggled chips unless the smuggling estimate is switched on, and Epoch describes that estimate as speculative.

What the US line means. Epoch gives no US total. "Seven US companies" is the sum of the owners it names individually (Amazon, CoreWeave, Google, Meta, Microsoft, Oracle and xAI), so it understates the US. "Everything outside China" adds Epoch's "Other" group, which mixes smaller US buyers with buyers in other countries. The lag is measured against the seven companies and is therefore a minimum.

Events are marked for reference only. A change in the trend after an export control does not show that the control caused it.

Best chip deployed at scale

The best chip each side actually owns in volume, quarter by quarter. A chip type counts once an owner group's cumulative units pass , using Epoch AI's chip-owner estimates through its last complete quarter.

Why bandwidth. Training and serving large models is usually limited by how fast a chip can move data to and from memory, which depends mostly on access to high-bandwidth memory (HBM), rather than by its peak arithmetic.

How sensitive this is to the threshold

Every chip type counted as deployed

Training compute

The largest training run with a published estimate, by country, in floating-point operations (log scale). Estimates by Epoch AI. Disclosure differs so much between the two countries that this is a comparison of what is disclosed, not of what was trained.

How many frontier models have an estimate

Capability against training compute

Each point is a model with both an Epoch Capabilities Index score and a training-compute estimate, released since 2023.

US modelChinese model

Chinese models often sit above US models at the same compute. Because the US points are limited to models that disclose, "China achieves more per unit of compute" is suggestive, not proven.

Chip designs on paper

Announced specifications, not deployed chips. Best 16-bit performance of any chip released to date (the higher of tensor-core FP16/BF16 and plain FP16), Nvidia against Chinese designers, log scale. A design made in small numbers counts the same as one made in millions, so this is not a measure of what either side can use.

Cost

What it costs to use a model of a given capability from each country, at today's list prices. The months-behind rule does not apply yet, because there is no price history to look back through.

Cheapest model at a given level

For each score on the Artificial Analysis Intelligence Index, the cheapest US and cheapest Chinese model that reaches it. Price is US dollars per million tokens, blended three parts input to one part output.

Price against capability

Each point is one model configuration with a list price. The stepped lines are each country's price frontier: the cheapest way to get at least that score.

These are list prices. They show what a user pays per token, not what it costs to run the model. Some providers price below cost, and a model that uses more tokens per task costs more than its per-token price suggests.

Policy timeline

US export controls, Entity List actions and Chinese chip milestones, with where the capability lag and the chip stock stood at the time of each.

Timing alone does not show cause. The lag moves for unrelated reasons: models are released in bursts, and benchmarks enter and leave the calculation. Read the marks as context, not as effects.

Methods and data

How every number on this site is calculated, what counts, and what is still uncertain.

One rule

Take China's best result released by a date. Find the earliest US result, released by that date, that was at least as good. The lag is the time between that US result and the date. If no US result is as good, China leads, and the same rule runs in reverse to give a negative lag.

Capabilities: the result is a benchmark score, dated by the model's release. An area's lag is the median across its counted benchmarks.
Chip stock: the result is cumulative H100-equivalents at each quarter end, so the lag moves in three-month steps and is measured to Epoch's last complete quarter.
Best deployed chip: the result is the per-chip 16-bit performance (or memory bandwidth) of the best chip type an owner group holds in volume, dated to the quarter its cumulative units passed the threshold.
Training compute is shown as the largest disclosed run by country, without a months-behind headline, because far fewer recent US models disclose compute.
Cost: the rule is not applied. Prices are known only for today, so the page compares them directly and records them daily to build a history.

What counts

Counted benchmarks: at least 3 Chinese models released in the last 12 months and one in the last 6, and not saturated or closed.
Saturated: both countries' best within 5% of the ceiling, or on a hand-kept list of saturated benchmarks.
Low confidence: an area with fewer than 3 counted benchmarks, or whose benchmarks disagree on direction. Shown faded, never hidden, and left out of headlines.
Published estimates: shown beside an area with fewer than 2 counted benchmarks, labelled with publisher, date and what they measure.
Independent sources: results not reported by the model's own developer. The toggle on the capabilities tab switches to all sources.
Before 2025: shaded on every time chart, because few Chinese models, and almost no closed ones, were scored.

Caveats specific to compute and cost

The China chip bucket is chips Epoch infers were sold to China-based companies, plus Chinese-designed chips. It is not government ownership and excludes smuggled chips unless that estimate is switched on.
The smuggling estimate is Epoch's own model output and Epoch calls it speculative. It starts in the first quarter of 2024.
No US total exists in the chip data. Seven named US companies stand in for the US, which understates it and makes the lag a minimum.
Deployed chips rest on Epoch's chip-owner estimates, which start in 2022 (2024 for Huawei and Cambricon) and give no US total. Chip types are matched to specifications by a hand-kept table; mismatches are flagged on the compute tab.
Per-chip figures understate system-level scaling, such as Huawei linking 384 Ascend chips in one system.
The spec-sheet chart is a specification, not a count of chips made. A chip that was announced but produced in small numbers counts the same as one produced in millions, so it is shown last and never on the overview.
Training compute has disclosure bias: most recent US closed models have no estimate.
Prices are list prices, not costs, and the history starts on the day this view was added.

Data to check

Rows carrying a flag or a note that something could not be confirmed.

Recent releases

Model	Country	Weights	ECI	Benchmarks	Released

Sources and attribution

Epoch AI: benchmark data, Epoch Capabilities Index, AI chip owners, ML hardware and model training compute. Licensed CC BY 4.0.
Arena leaderboard dataset by LMArena. Licensed CC BY 4.0. Measures human preference, not capability.
METR time horizons, a closed series.
Events and published estimates are linked to their sources row by row.